Research On The Theory And Application Of Deep Learning Optimization

Neural networks have demonstrated outstanding efficacy in diverse domains,particularly in image classification tasks,and have thus been extensively applied in recent years.The effectiveness of neural networks heavily relies on the choice of their hyperparameters,and different model hyperparameters have different performance effects.The characteristics of hyperparametr optimization problems in neural networks are large computational complexity,interdependence and wide range of related parameters,and complex network architecture design.Therefore,for traditional manual parameter adjustment,the workload is huge and the calculation time is long,and the final results are highly dependent on the level and experience of the parameter adjustment designer.Swarm intelligence algorithms have received widespread attention from researchers in various fields due to their powerful capabilities,high flexibility,and ease of parallelism.Therefore,this paper focuses on intelligent optimization algorithms,studies neural network hyperparametric optimization problems,and applies them to image classification problems.The main work of this article is as follows:(1)For the optimization of hyperparameters with a given neural network model,a globally optimal dimensionally adjusted Artificial Bee Colony algorithm(GDABC)is proposed and applied to the diagnosis of Parkinson’s disease.The GDABC algorithm is an enhanced version of the traditional Artificial Bee Colony(ABC)algorithm used to optimize the hyperparameters of deep learning models such as neural networks.It proposes three modifications to improve the ABC algorithm’s performance,including a hybrid coding strategy that maps hyperparameters to continuous domains,a range pruning strategy that speeds up the convergence of the algorithm,and a dimension adjustment strategy that enhances the local mining ability of the algorithm.The algorithm has been tested on the MNIST dataset to optimize neural network models and has shown promising results in improving the performance of machine learning models.Finally,aiming at the problem that Parkinson’s disease is difficult to diagnose in the early stage,a Parkinson’s auxiliary diagnosis system based on deep learning hyperparametric optimization is proposed using this algorithm.Experiments show that the diagnostic accuracy of the Res Net50 network optimized by the GDABC algorithm is higher than the current mainstream optimization methods and Parkinson’s diagnosis methods.(2)For the optimization of hyperparameters with a given neural network model,an Intelligent Bee Colony Algorithm for Intelligent Search Optimization(ISABC)was proposed and applied to face recognition with age invariance.To avoid local optima when optimizing hyperparameters,the ISABC algorithm utilizes a reverse learning strategy for reconnaissance bees which helps the algorithm explore new areas in the hyperparameter space and improve its overall performance by avoiding becoming stuck in suboptimal regions.To improve the mining ability of the ISABC algorithm and overcome the limited local search ability of GDABC,an intra circle dance strategy is added to the dimension adjustment strategy which enhances the algorithm’s ability to find optimal solutions in the hyperparameter space.Experiments show that applying the ISABC algorithm to optimize hyperparameters of a neural network model leads to better performance on CIFAR10 image classification dataset,demonstrating the algorithm’s effectiveness.Finally,for age invariant face recognition,the algorithm is used to optimize the network’s hyperparameters.Experiments show that the Res Net18 network optimized by the ISABC algorithm improves the accuracy of face recognition.(3)For the optimization of hyperparameters without a given neural network model(the design of neural network structures),an Artificial Bee Colony Algorithm Based on Network Structural Search(NASABC)was proposed and applied to human motion recognition.The algorithm designs a coding strategy based on the artificial bee colony algorithm to achieve network structure coding,and proposes a new update strategy to achieve network structure search.Finally,on the UCI-HAR dataset,the NASABC algorithm is used to automatically design the structure of the network and achieve human motion behavior recognition.Based on the characteristics of high flexibility and easy parallelization of the artificial bee colony algorithm,this paper proposes three intelligent algorithms for optimizing the hyperparameter of neural networks,and uses the three intelligent algorithms to solve the actual classification problem.This research not only provides a solution for the optimization of neural network hyperparameter,but also provides a new idea for improving the classification accuracy.